Face Recognition in Ideal and Noisy Conditions Using Support Vector Machines, PCA and LDA
Miloš Oravec, Jan Mazanec, Jarmila Pavlovičová, Pavel Eiben, Fedor Lehocki · InTech eBooks · 2010
We examined different scenarios of face recognition experiments. They contain both singlestage and two-stage recognition systems. Single-stage face recognition uses SVM for classification directly. Two-stage recognition systems include PCA with MahCosine metric, LDA with LDASoft metric and also methods utilizing both PCA and LDA for feature extraction followed by SVM for classification. All methods are significantly influenced by different settings of parameters that are related to the algorithm used (i.e. PCA, LDA or SVM). This is the reason we presented serious analysis and proposal of parameter settings for the best performance of discussed methods. For methods working in ideal conditions, the conclusions are as follows: When comparing non-SVM based methods, higher maximum recognition rate is generally achieved by method LDA+LDASoft compared to PCA+MahCosine; on the other hand LDA+LDASoft is more sensitive to method settings. Using SVM in classification stage (PCA+SVM and LDA+SVM) produced better maximum recognition rate than standard PCA and LDA methods. Experiments with single-stage SVM show that this method is very efficient for face recognition even without previous feature extraction. With 4 images per subject in training set, we reached 96.7% recognition rate. The experiments were made with complex image set selected from FERET database containing 665 images. Such number of face images entitles us to speak about general behavior of presented methods. Altogether more than 600 tests were made and maximum recognition rates near 100% were achieved. It is important to mention that the experiments were made with "closed" image set, so we did not have to deal with issues like detecting people who are not in the training set. On the other hand, we worked with real-world face images; our database contains images of the same subjects that often differ in face expressions (smiling, bored, ...), with different hairstyles, with or without beard, or wearing glasses and that were taken in different session after longer time period (i.e. we did not work with identity card-like images).